You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@cassandra.apache.org by William Katsak <wk...@cs.rutgers.edu> on 2014/10/08 18:27:26 UTC
Consistency Levels
Hello,
I was wondering if anyone (Datastax?) has any usage data about
consistency levels. For example, what consistency levels are real
applications using in real production scenarios. Who is using eventual
consistency (ONE-ONE) in production vs strong consistency
(QUORUM-QUORUM, ONE-ALL). Obviously it depends on the application, but I
am trying to collect some information on this.
I saw the talk from Christos Kalantzis (from Cassandra13 I think) about
Netflix using eventual consistency, but I was wondering if there is any
more data out there.
Thanks in advance,
Bill Katask
Ph.D. Student
Department of Computer Science
Rutgers University
Re: Consistency Levels
Posted by DuyHai Doan <do...@gmail.com>.
One should be carefull about using ALL consistency because by doing so, you
sacrify the high availability (loosing one node of the replica prevent you
from writing/reading with ALL). Lots of people choose Cassandra for high
availability so using ALL is kind of showstopper.
Of course there are specific cases where such level can be relevant but
generally I advise people to use the couple QUORUM/QUORUM rather than
ONE/ALL or ALL/ONE if they want stronger consistency than ONE/ONE. The
latter combination is used for low latency when immediate consistency is
not a requirement. Users rely on all the anti-entropy processses
(read-repair, consistent read, scheduled repair) to make data converge.
On Wed, Oct 8, 2014 at 6:27 PM, William Katsak <wk...@cs.rutgers.edu>
wrote:
> Hello,
>
> I was wondering if anyone (Datastax?) has any usage data about consistency
> levels. For example, what consistency levels are real applications using in
> real production scenarios. Who is using eventual consistency (ONE-ONE) in
> production vs strong consistency (QUORUM-QUORUM, ONE-ALL). Obviously it
> depends on the application, but I am trying to collect some information on
> this.
>
> I saw the talk from Christos Kalantzis (from Cassandra13 I think) about
> Netflix using eventual consistency, but I was wondering if there is any
> more data out there.
>
> Thanks in advance,
>
> Bill Katask
> Ph.D. Student
> Department of Computer Science
> Rutgers University
>
Re: Consistency Levels
Posted by Jack Krupansky <ja...@basetechnology.com>.
It would certainly depend on the nuances of your definitions! Especially
since you added this monster of a caveat: " in the absence of failures".
Here's a scenario for you: 1) Write at quorum, 2) Add 3 nodes, 3)
Immediately read at [the new] quorum - no guarantee that the new nodes will
have fully bootstrapped.
Who knows how many other modalities there might be - despite however many
caveats you want to tack on.
"Strong consistency" is a "model", not necessarily a reality at any point in
time even if it was a reality at a prior point in time.
If I deliberately "decommission" a node, that isn't necessarily a "failure"
is it?
All of that said, "it depends" on where you're trying to get to.
-- Jack Krupansky
-----Original Message-----
From: William Katsak
Sent: Wednesday, October 8, 2014 7:19 PM
To: user@cassandra.apache.org
Subject: Re: Consistency Levels
Thanks.
I am thinking more in terms of the combination of read/write. If I am
correct, QUORUM reads and QUORUM writes (or ONE-ALL) should deliver
strong consistency in the absence of failures, correct? Or this this
still considered eventual consistency, somehow?
-Bill
On 10/08/2014 06:17 PM, Jack Krupansky wrote:
> I don't know of any such data collected by DataStax - it's not like
> we're the NSA, sniffing all requests.
>
> ONE is certainly fast, but only fine if you don't have immediate need to
> read the data or don't need the absolutely most recent value.
>
> To be clear, even QUORUM write is eventual consistency - to all nodes
> beyond the immediate quorum.
>
> -- Jack Krupansky
>
> -----Original Message----- From: William Katsak
> Sent: Wednesday, October 8, 2014 12:27 PM
> To: user@cassandra.apache.org
> Subject: Consistency Levels
>
> Hello,
>
> I was wondering if anyone (Datastax?) has any usage data about
> consistency levels. For example, what consistency levels are real
> applications using in real production scenarios. Who is using eventual
> consistency (ONE-ONE) in production vs strong consistency
> (QUORUM-QUORUM, ONE-ALL). Obviously it depends on the application, but I
> am trying to collect some information on this.
>
> I saw the talk from Christos Kalantzis (from Cassandra13 I think) about
> Netflix using eventual consistency, but I was wondering if there is any
> more data out there.
>
> Thanks in advance,
>
> Bill Katask
> Ph.D. Student
> Department of Computer Science
> Rutgers University
--
****************************************
William Katsak <wk...@cs.rutgers.edu>
Ph.D. Student
Rutgers University
Department of Computer Science
****************************************
Re: Consistency Levels
Posted by William Katsak <wk...@cs.rutgers.edu>.
Thanks.
I am thinking more in terms of the combination of read/write. If I am
correct, QUORUM reads and QUORUM writes (or ONE-ALL) should deliver
strong consistency in the absence of failures, correct? Or this this
still considered eventual consistency, somehow?
-Bill
On 10/08/2014 06:17 PM, Jack Krupansky wrote:
> I don't know of any such data collected by DataStax - it's not like
> we're the NSA, sniffing all requests.
>
> ONE is certainly fast, but only fine if you don't have immediate need to
> read the data or don't need the absolutely most recent value.
>
> To be clear, even QUORUM write is eventual consistency - to all nodes
> beyond the immediate quorum.
>
> -- Jack Krupansky
>
> -----Original Message----- From: William Katsak
> Sent: Wednesday, October 8, 2014 12:27 PM
> To: user@cassandra.apache.org
> Subject: Consistency Levels
>
> Hello,
>
> I was wondering if anyone (Datastax?) has any usage data about
> consistency levels. For example, what consistency levels are real
> applications using in real production scenarios. Who is using eventual
> consistency (ONE-ONE) in production vs strong consistency
> (QUORUM-QUORUM, ONE-ALL). Obviously it depends on the application, but I
> am trying to collect some information on this.
>
> I saw the talk from Christos Kalantzis (from Cassandra13 I think) about
> Netflix using eventual consistency, but I was wondering if there is any
> more data out there.
>
> Thanks in advance,
>
> Bill Katask
> Ph.D. Student
> Department of Computer Science
> Rutgers University
--
****************************************
William Katsak <wk...@cs.rutgers.edu>
Ph.D. Student
Rutgers University
Department of Computer Science
****************************************
Re: Consistency Levels
Posted by Jack Krupansky <ja...@basetechnology.com>.
I don't know of any such data collected by DataStax - it's not like we're
the NSA, sniffing all requests.
ONE is certainly fast, but only fine if you don't have immediate need to
read the data or don't need the absolutely most recent value.
To be clear, even QUORUM write is eventual consistency - to all nodes beyond
the immediate quorum.
-- Jack Krupansky
-----Original Message-----
From: William Katsak
Sent: Wednesday, October 8, 2014 12:27 PM
To: user@cassandra.apache.org
Subject: Consistency Levels
Hello,
I was wondering if anyone (Datastax?) has any usage data about
consistency levels. For example, what consistency levels are real
applications using in real production scenarios. Who is using eventual
consistency (ONE-ONE) in production vs strong consistency
(QUORUM-QUORUM, ONE-ALL). Obviously it depends on the application, but I
am trying to collect some information on this.
I saw the talk from Christos Kalantzis (from Cassandra13 I think) about
Netflix using eventual consistency, but I was wondering if there is any
more data out there.
Thanks in advance,
Bill Katask
Ph.D. Student
Department of Computer Science
Rutgers University
Re: Consistency Levels
Posted by Robert Coli <rc...@eventbrite.com>.
On Wed, Oct 8, 2014 at 9:27 AM, William Katsak <wk...@cs.rutgers.edu>
wrote:
> I was wondering if anyone (Datastax?) has any usage data about consistency
> levels. For example, what consistency levels are real applications using in
> real production scenarios. Who is using eventual consistency (ONE-ONE) in
> production vs strong consistency (QUORUM-QUORUM, ONE-ALL). Obviously it
> depends on the application, but I am trying to collect some information on
> this.
>
Anecdotally, my sense of the typical deploy is that it is RF=N=3, with
CL.ONE.
=Rob